Could AI Make Drug Patents Obsolete?

Danaher announced this week that it is building an AI-powered autonomous lab at Abcam, which by itself probably does not mean much anymore. We’ve been hearing about autonomous labs for a while now, and my guess probably a lot more in the coming years.

The idea is basically that AI designs an experiment, robots run it, the results go back into the system, and then the AI uses those results to decide what to try next. Danaher says its system could eventually make antibody discovery up to eight times faster and generate ten times as many reagents every year.

But that’s not what I want to focus on this week, there’s another aspect to this story, or rather these stories that will be a big part of the conversation, which is patents.

So, just a bit of background so we’re all on the same page, a patent is basically an agreement between an inventor and the government. You disclose how your invention works and, in exchange, you generally get the exclusive right to that invention for a limited period of time. For most US utility patents, that is around 20 years from filing.

There is another important requirement though. Someone actually has to be the inventor, and right now, in the United States, that someone has to be a person, NOT an AI.

This has already gone through the courts. A computer scientist named Stephen Thaler tried to patent inventions created by an AI system called DABUS and listed the AI itself as the inventor. The US Patent Office rejected it, the Federal Circuit agreed, and the Supreme Court eventually declined to hear the appeal. Under current US patent law, an inventor has to be a “natural person.”

The USPTO reinforced this again in 2025. AI can absolutely be used in the invention process, but it is treated as a tool. Ultimately, there still needs to be a human contribution to the conception of the invention

OK, now back to our Danaher autonomous lab story, and how that could play out.

A scientist might tell the system that they want an antibody against a particular target with a certain affinity, stability and specificity. The AI generates thousands of sequences, robots make them, the lab tests them, the results go back into the model and the whole thing repeats.

Maybe after 15 or 20 rounds, antibody #4321 turns out to be fantastic, and is exactly what they’re looking for. Great - but who invented antibody #4321? I actually don’t know what the right answer to that is.

Was it the scientist who defined the target? The people who built the model? The scientist who chose some of the constraints? Or does someone eventually have to look at antibody, make some meaningful change to it, and essentially create the “human” contribution needed for the patent?

I think that possibility is pretty interesting because we could end up in this odd situation where the science becomes increasingly autonomous, but we deliberately keep a human somewhere in the processes because our patent system still needs someone to have invented the thing.

And maybe that works perfectly well. AI has been involved in inventions for years and the law has adapted to new tools before, but I wonder how far that logic will go before the whole system needs a complete overhaul.

Patents were created in a world where invention was a relatively scarce human activity, it is a very difficult endeavour and should be encouraged. So giving someone temporary exclusivity was one way of getting people to invent and publicly disclose what they had made. An autonomous lab could eventually generate thousands or even millions of plausible inventions continuously.

My point here is, what happens to a patent system built around a HUMAN invention that no longer works. Is it even worth patenting because with the right resource, another AI model could ‘invent’ the same thing or even reverse engineer the patent and produce an even better version. Something to think about!

Iambic just submitted an application to the FDA to begin human trials of its second AI-discovered cancer drug. The drug, called IAM217, targets a protein involved in cell division and is being developed for breast, ovarian and other solid tumours. What's particularly interesting is that Iambic reported around 90% tumour regression in a preclinical brain tumour model, which is promising considering how difficult it can be to get drugs into the brain. Of course, these are still animal results, so we'll have to see how it performs in humans. [Link]

Geoffrey Hinton, often called the "Godfather of AI," thinks AI companies should have to go through something similar to FDA approval before releasing their models. His argument is that pharmaceutical companies can't just develop a new drug and immediately sell it to the public. They first have to demonstrate that it's safe and effective to regulators, a process that can take years and cost enormous amounts of money. Hinton thinks AI developers should have to demonstrate safety before release as well, which is an interesting comparison given how quickly new AI models are coming out. [Link]

This year's Nobel Prize in Chemistry went to Henri Kagan and Kenso Soai for their work on molecular handedness, something that's particularly important in drug development. Basically, many molecules can exist in two mirror-image forms, kind of like your left and right hands. Even though they contain the same atoms, our bodies can respond very differently to each version. Their discoveries helped chemists understand and control how one form can be produced over the other, which has become an important part of modern pharmaceutical chemistry. [Link]